Advances in approximate Bayesian computation and trans-dimensional sampling methodology

Abstract

Bayesian statistical models continue to grow in complexity, driven in part by a few key factors: the massive computational resources now available to statisticians; the substantial gains made in sampling methodology and algorithms such as Markov chain Monte Carlo (MCMC), trans-dimensional MCMC (TDMCMC), sequential Monte Carlo (SMC), adaptive algorithms and stochastic approximation methods and approximate Bayesian computation (ABC); and development of more realistic models for real world phenomena as demonstrated in this thesis for financial models and telecommunications engineering. Sophisticated statistical models are increasingly proposed for practical solutions to real world problems in order to better capture salient features of increasingly more complex data. With sophistication comes a parallel requirement for more advanced and automated statistical computational methodologies.

The key focus of this thesis revolves around innovation related to the following three significant Bayesian research questions. 1. How can one develop practically useful Bayesian models and corresponding computationally efficient sampling methodology, when the likelihood model is intractable? 2. How can one develop methodology in order to automate Markov chain Monte Carlo sampling approaches to efficiently explore the support of a posterior distribution, defined across multiple Bayesian statistical models? 3. How can these sophisticated Bayesian modelling frameworks and sampling methodologies be utilized to solve practically relevant and important problems in the research fields of financial risk modeling and telecommunications engineering ?

This thesis is split into three bodies of work represented in three parts. Each part contains journal papers with novel statistical model and sampling methodological development. The coherent link between each part involves the novel sampling methodologies developed in Part I and utilized in Part II and Part III. Papers contained in each part make progress at addressing the core research questions posed.

Part I of this thesis presents generally applicable key statistical sampling methodologies that will be utilized and extended in the subsequent two parts. In particular it presents novel developments in statistical methodology pertaining to likelihood-free or ABC and TDMCMC methodology. The TDMCMC methodology focuses on several aspects of automation in the between model proposal construction, including approximation of the optimal between model proposal kernel via a conditional path sampling density estimator. Then this methodology is explored for several novel Bayesian model selection applications including cointegrated vector autoregressions (CVAR) models and mixture models in which there is an unknown number of mixture components. The second area relates to development of ABC methodology with particular focus on SMC Samplers methodology in an ABC context via Partial Rejection Control (PRC). In addition to novel algorithmic development, key theoretical properties are also studied for the classes of algorithms developed. Then this methodology is developed for a highly challenging practically significant application relating to multivariate Bayesian $\alpha$-stable models.

Then Part II focuses on novel statistical model development in the areas of financial risk and non-life insurance claims reserving. In each of the papers in this part the focus is on two aspects: foremost the development of novel statistical models to improve the modeling of risk and insurance; and then the associated problem of how to fit and sample from such statistical models efficiently. In particular novel statistical models are developed for Operational Risk (OpRisk) under a Loss Distributional Approach (LDA) and for claims reserving in Actuarial non-life insurance modelling. In each case the models developed include an additional level of complexity which adds flexibility to the model in order to better capture salient features observed in real data. The consequence of the additional complexity comes at the cost that standard fitting and sampling methodologies are generally not applicable, as a result one is required to develop and apply the methodology from Part I.

Part III focuses on novel statistical model development in the area of statistical signal processing for wireless communications engineering. Statistical models will be developed or extended for two general classes of wireless communications problem: the first relates to detection of transmitted symbols and joint channel estimation in Multiple Input Multiple Output (MIMO) systems coupled with Orthogonal Frequency Division Multiplexing (OFDM); the second relates to co-operative wireless communications relay systems in which the key focus is on detection of transmitted symbols. Both these areas will require advanced sampling methodology developed in Part I to find solutions to these real world engineering problems.